On design of nonlinear robotic control system with neural networks

摘要

The authors explore the use of neural networks in conjunction with a learning control scheme to simplify the mathematical complexity in nonlinear robotic controls. A sufficient-input requirement is proposed for neural network applications to dynamic systems based on the system invertibility theory. A lower bound is found which determines the least number of inputs required to excite a neural network designed to learn the inverse of a robotic system. In the proposed combination method of robotic learning control, a nonlinear learning law is first applied on a robotic system to find a desired basin of the output error. Then, using a back-propagation neural network (BPN) to learn the system input-output relation from previously recorded data, an expected control input with respect to the desired task for the robotic system can iteratively be resolved by a learning process. A simulation study has been carried out to verify the convergence of the combination learning method associated with a BPN. From the simulation results, convergence performance has been achieved for the Stanford-like robot arm.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>